Google AlphaEvolve Goes GA: Gemini-Powered Code Optimization Agent Now Available

Google this week announced that AlphaEvolve, evolved from a DeepMind research project, is now Generally Available on the Gemini Enterprise Agent Platform. The service is an evolutionary code optimization and algorithm-discovery agent that uses Gemini models to mutate baseline code and search algorithmic space for highly optimized, human-readable versions.

AlphaEvolve operates through a strict four-step workflow: users provide a seed algorithm, build a deterministic client-side evaluation script to score candidates on performance and correctness, the AI generates and searches mutated candidates, and finally the winning code is deployed. Crucially, code generation happens through Google's API while the evaluator script runs client-side on the user's own infrastructure, dividing labor to protect data.

However, the tool has clear limitations. AlphaEvolve is completely ineffective for standard business logic with ambiguous requirements — the outcome must be mathematically scorable. If the evaluator script misses a constraint, the evolutionary search will ruthlessly exploit it, generating code that passes flawed tests but fails catastrophically on edge cases. All published metrics remain vendor-supplied with no independent third-party benchmarks yet.

In biosciences, AlphaEvolve has shown remarkable results. The Qin Lab at Old Dominion University applied it to computational biogerontology, independently rediscovering the 1990s Kannisto logistic mortality model in just 500 evaluations and improving the Emergent Aging Model composite fitness score by 19%. PacBio used it on Google's DeepConsensus model to slash DNA sequencing variant detection errors by 30%. Schrödinger quadrupled machine learning force field inference speed, drastically shortening R&D cycles for molecular simulations in drug discovery.

Enterprise customers are reporting significant gains. Klarna doubled its ML training throughput while exploring nearly 6,000 deep architectural rewrites in three weeks. JetBrains reduced IDE code completion latency by 15-20%. FM Logistic cut warehouse picking routing costs by 10.4%. Kinaxis boosted supply-chain forecast accuracy by 22% while dropping runtime by 90%.

Internally, Google used AlphaEvolve to design next-generation TPU silicon layouts, reduced Google Spanner's write amplification by 20%, cut software storage footprints by 9%, and discovered Willow quantum processor circuits with 10x lower error rates.

Developers can access the AlphaEvolve API via documentation and Colab examples, or use the AlphaEvolve Skill to integrate the optimization workflow directly into IDEs like Antigravity and Claude Code. An open-source alternative called OpenEvolve is also available.